NVIDIA's Jensen Huang: Criticizes AI Doomsday Theory
NVIDIA CEO Jensen Huang criticized the prediction that AI could lead to human extinction as "made up" and "irresponsible" during an interview at the "All-In" summit on September 14. He specifically pointed out that the claim by some researchers that the probability of AI causing human extinction in the next decade exceeds 10% has "no scientific basis" and stated that "such predictions should not have been made." Huang then listed a series of AI predictions he believes have been proven wrong in the past, including that radiology would be fully automated with no need for radiologists, that 90% of code would be generated by AI within 6 to 12 months, and that 50% of entry-level jobs would disappear within 6 to 9 months. He emphasized that "all these predictions have ultimately been proven wrong" and stressed that those who make predictions should be held accountable for their inaccuracies.
Huang's comments were directly prompted by a resignation statement from 27-year-old AI researcher Jacob Coxon at Anthropic, who had previously worked on GPT-4 at OpenAI and left Anthropic after just four months, willing to forfeit unvested equity. Coxon claimed the industry is "rushing towards self-improving superintelligence, betting human lives" and asserted that those developing AI genuinely believe this technology could destroy humanity by the end of the century. Anthropic's alignment science lead Evan Hubinger later publicly endorsed this view, revealing that he assessed the probability of AI causing human extinction in the next decade to exceed 10%, which Huang referred to as a "made up" prediction.
The debate has quickly spread to the upper echelons of the AI industry. Anthropic CEO Dario Amodei recently published an article titled "We Must Pace the Frontier," calling for a slowdown in AI development, a stance echoed by some executives at OpenAI, Microsoft, and Google. Huang, in an interview with Axios on September 10, attributed the discussion to commercial motives, stating that "the cybersecurity industry is talking about this topic everywhere because it is preparing to launch new products... what better way to create demand than to manufacture a problem?" He also refuted Bill Gates' earlier warning that AI would lead to mass unemployment.
At the Salesforce Dreamforce conference on September 15, Huang further elaborated on his regulatory stance, stating that "safety is an engineering issue, not a legal issue... we do not need new laws, we do not need new regulations," advocating that "innovation, speed, and safety products" are not mutually exclusive. He also proposed that independent third-party "assessment agencies" take on roles similar to financial auditors—needing not to be more professional than the companies themselves but capable of "asking the right questions," and that multiple assessment agencies should be introduced to avoid a single company dominating the narrative.
The underlying commercial interests are clear—Huang, as the largest supplier of AI computing infrastructure globally, has a core business interest in the continuous acceleration of AI development and deployment, as well as expanding capital expenditures. Therefore, he is naturally inclined to downplay extinction-level risk narratives and oppose new regulations to avoid a slowdown in demand for NVIDIA chips due to "pauses" or rising compliance costs. He himself acknowledged that NVIDIA's investments in downstream AI companies yield "one dollar invested returns one hundred dollars," highlighting the intertwined interests throughout the AI supply chain. In contrast, the warnings from Anthropic and its internal safety researchers objectively reinforce the narrative of "the need for caution, assessment, and investment in safety resources," intersecting with the cybersecurity industry that provides AI safety audits, red team testing, and safety products—this is precisely the motive Huang emphasized in his Axios interview. The beneficiaries are the computing power and model providers advocating for "acceleration priority"; the pressured parties are the safety research camp and related security service providers pushing for a slowdown, whose credibility is now under scrutiny in the market.
Notably, Huang did not completely dismiss the safety issue itself in the interview—he publicly praised Coxon for "showing tremendous courage" and stated that if companies "believe they are out of control," they should proactively slow down their development pace, while also acknowledging that the path to "superintelligence" is "reasonable," but the related expressions are being "weaponized," and companies still need to ensure product safety through "control, verification, and assessment" before release.
Source: Public Information
ABAB AI Insight
Huang's public questioning of the technology panic narrative is not an isolated case—since the generative AI wave triggered by ChatGPT in 2022, NVIDIA has repeatedly publicly promoted its AI chips despite concerns about "overcapacity" and has continuously injected funds into downstream AI companies like OpenAI, xAI, and CoreWeave through equity investments and cloud computing purchase agreements, which then purchase NVIDIA chips, forming what is referred to as a "circular financing" model. Huang has previously defended this model, stating that "investing one dollar can yield one hundred dollars," and his recent rebuttal of the "AI extinction theory" continues his consistent style of "technological optimism + defending capital expenditure narrative."
From a funding chain perspective, NVIDIA's business model heavily relies on the continuous expansion of AI capital expenditures—through direct equity investments and cloud computing purchase agreements, NVIDIA injects funds into multiple AI startups, which are primarily used to purchase NVIDIA GPUs, forming the closed loop Huang refers to as "invest one, return one hundred"; if the "AI extinction theory" or large-scale regulatory narratives dominate mainstream discourse, it will likely directly increase compliance and insurance costs for data center construction and may trigger a revaluation of the sustainability of AI capital expenditures in the capital market—this is precisely the capital motivation behind Huang's willingness to publicly oppose Anthropic and some safety researchers and insist on the stance of "no new regulations."
This is highly reminiscent of historical narratives where "new technology practitioners refute safety warnings"—a more direct analogy is the nuclear energy industry: nuclear power practitioners have long downplayed systemic risk narratives by arguing that "Chernobyl and Fukushima are extreme cases," while opponents have continuously used "low probability, high destructiveness" events as a basis for regulation; Huang's examples of "radiology automation failures" and "programming automation predictions falling short" essentially use "past prediction inaccuracies" to counter the argument that "this time might be different." The current AI industry is at a critical window of "capital expenditure explosion and safety regulatory game": on one side, manufacturers like NVIDIA are pushing for "building while verifying," while on the other side, internal researchers at Anthropic are issuing warnings through resignations, shifting the industry discourse from a purely technical competition to a struggle for safety governance dominance.
This is fundamentally a struggle for pricing power within the AI industry chain due to differences in business models—NVIDIA, as the "shovel seller," has revenue directly tied to the expansion of computing power demand, naturally inclined to characterize safety disputes as "artificially created demand traps" rather than real engineering risks; model vendors like Anthropic and their internal safety teams, possessing firsthand information about the boundaries of model capabilities, have the narrative dominance to package "slowing down" as industry consensus. If the narrative of "the need for cautious assessment" prevails, regulatory costs and assessment processes themselves may evolve into new barriers to entry and business opportunities (such as the third-party "assessment agency" model Huang proposed). Mechanistically, as long as the real risk level of AI capabilities remains in a state of information asymmetry—outsiders cannot independently verify whether models truly possess the ability for "self-improvement" or "deceptive testing," both sides have motives to interpret uncertainty in the version most favorable to their business models. This debate over whether the "extinction theory is made up" is essentially a contest over who can define AI safety standards and thus control the dominant narrative of entry and compliance in the industry.
ABAB News · Cognitive Law
- The shovel seller never believes there is a risk in gold mining.
- Whoever controls the definition of panic controls the entry threshold.
- Losing predictions doesn't matter, as long as the business wins.